动态软测量技术在食品发酵过程中的应用

Application of dynamic soft sensing technology in food fermentation process

中文关键词:  发酵  软测量  局部加权偏最小二乘  滑动时间窗  海洋碱性蛋白酶

英文关键词:fermentation  soft sensor  locally weighted parital least squares  moving window  marine alkaline protease

中文摘要:

      目的 针对食品发酵生产过程中关键参数难以在线检测, 不能对其实施优化控制的问题。本研究提出一种即时学习和集成学习相结合的动态软测量建模方法。方法 首先采用局部加权偏最小二乘算法(locally weighted parital least squares, LWPLS)作为基础建模方法, 再运用滑动时间窗技术(moving window, MW)缩小样本查询范围。采用Bagging算法重采样窗口内的样本集形成多个子训练集, 同时引入相似度阈值筛选出具有多样性的子训练集, 最后采用加权平均法融合各个子模型的输出值。结果 通过海洋碱性蛋白酶发酵实验表明, 所提的MW-ELWPLS软测量建模方法能够对关键参数进行较准确的实时预测, 预测的均方根误差在0.2~0.3之间, 决定系数R2在0.95左右, 平均绝对误差在0.2左右。结论 MW-ELWPLS软测量建模方法在预测精度上表现优秀, 在线预测实时性较高, 完全能满足工业发酵生产的实际应用需求。

英文摘要:

      Objective To propose a dynamic soft sensor modeling method combining just in time learning and ensemble learning to solve the problem that the key parameters in food fermentation production are difficult to be detected online, let alone to be controlled optimally. Methods Local weighted partial least squares (LWPLS) was used as the basic modeling method, and moving window technology (MW) was used to narrow the sample query range. Then Bagging algorithm was used to resample the sample set in the window to form multiple sub-training sets. At the same time, similarity threshold was introduced to screen out diverse sub-training sets. Finally, weighted average method was adopted to fuse the output values of each sub-model. Results The marine alkaline protease fermentation experiment showed that the proposed MW-ELWPLS soft sensing modeling method could accurately predict the key parameters in real time. The predicted root mean square error was between 0.2?0.3, the determination coefficient R2 was about 0.95, and the mean absolute error was about 0.2. Conclusion MW-ELWPLS soft measurement modeling method has excellent prediction accuracy and high real-time performance in online prediction, which can fully meet the practical application needs of industry fermentation production.